FUTO

people love @FUTO_Tech t-shirts

A few weeks ago, we gathered in Austin to celebrate the progress of PALLADIUM and raise support for the work ahead. Thank you to all of our lovely friends and supporters for another hit, and a special thanks to @FUTO_Tech for sponsoring this event!

We have created the most optimal Swipe layout.

lIIIIlllllIIIIIIIIIIIlllllIIIIIIIIIlllllll Who thought this was a good idea? Why do we keep building it into all of our UI? Designers can you defend this?

I've started noticing major bugs and breakages in a bunch of software I use Like completely broken UI layouts and embarrassingly bad performance And it all seems to line up with the surge in AI generated code this year Just sayin

We just put up our preprint for the new FUTO Swipe model on arXiv, if you are interested in a more technical overview of it, definitely give it a read.

Want to take a deeper look into our new Notes app? Well, we've got just the video for you!

Huge thanks to @TheFreedomFest that have, along with @FUTO_Tech and @rossmannsupply, brought us to this precipice of technological freedom in mobility equipment.

Sneak preview our new app:

Closing the curtains doesn't require an explanation. Neither should using your phone. phreeli.com

@awesomekling @ladybirdbrowser

The Ladybird: State of the Union talk at our Don't Be Evil Conference is now available to watch right here!

@sugumaran___ We are posting a preprint on arXiv soon that goes into much greater detail, we will link it when its uploaded.

@FUTO_Tech FUTO Swipe Architecture: The system uses three models working together: Encoder: A universal, layout/language-agnostic model for general swipe typing, but not state-of-the-art. ContextLM: A tiny language model (trained per language on text only) that filters out nonsensical word options based on sentence context. Decoder: A layout- and language-specific model (currently only QWERTY English) that delivers top accuracy by learning layout details, though it requires swipe training data. Combined with a beam width of 300, the system achieves a ~4% top-4 failure rate (under 1% when ignoring out-of-vocabulary cases).

FUTO Swipe running a swipe keyboard model fully on-device is more interesting than it sounds. The real story isn't the typing. It's that someone shipped a sequence prediction model small enough to run in a keyboard process, with latency under the threshold where users notice. That's a harder engineering problem than most production ML I see in enterprise. Most teams I work with are still defaulting to API calls for anything that feels like 'AI', because on-device inference feels risky and hard to update. But the tradeoff is real. A keyboard sees everything you type. Sending that to a server is a privacy decision, not just an architecture decision. Under GDPR, that distinction matters in ways that US teams genuinely haven't had to think through yet. The FUTO approach forces the model to be small, fast, and updatable without a network hop. Those constraints produce better engineering than open-ended cloud budgets do. My students always want to build the biggest model. I keep telling them: fit your model to the constraint first. This is what that looks like in practice.

@ZeimusCS It is definitely in the nightly build as a toggle, not sure if it is in the version on Google Play Store.

@FUTO_Tech can I do two finger swyping with FUTO swipe?

This doesn't seem to affect the hidden agenda software developed by the biggest Californian companies. I'm sure Google, Apple, and FB are happy to punish any developers who want an honest relationship with their customers.